Inkling 2 Mini

Thinking Machines

Multimodal autoregressive looped transformer with sparse MoE feed-forward backbone

Open Weights Commercial Use Multimodal NVFP4, MXFP8 Inkling English Multilingual (general capabilities)

Parameters

26.2B / 5.4B active

MoE total / active

Context

1,048,576 tokens

native

Modalities

text, image text

input → output

Benchmarks

#1/138

better than 99% of all scored models

Size class

#1/5

better than 80% of similar-size models

License

Apache License 2.0

About

Inkling 2 Mini is a general-purpose multimodal open-weights model from Thinking Machines Lab, Inc. (release: November 20, 2026, license: Apache 2.0, model family Inkling, alongside Inkling 2 Small). This record refers to the official weights distribution on Hugging Face, thinkingmachines/Inkling-2-Mini; the checkpoint ships as a single ~20 GB file in mixed NVFP4 and MXFP8 precisions. Unsloth publishes quantized GGUF derivatives in the unsloth/Inkling-2-Mini-GGUF repository (Dynamic 3.0 / imatrix UD-* quants from UD-IQ1_S at 7.76 GB up to UD-Q8_K_XL at 22.87 GB, with a per-quant analysis of Mean KLD vs UD-Q8_K_XL and same-top-token rate, mainline llama.cpp uploads under Mainline_Support, and Unsloth Desktop support; the card of that GGUF repo was the source document of this ingest).

Model

Inkling 2 Mini accepts text and image inputs and generates text outputs. It is intended for English and other languages and across multiple coding languages, and for developers building AI-powered applications: agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation, plus general-purpose conversational use and instruction following. It supports a context window of up to 1M tokens. Images are encoded by a vision transformer (inkling2_vision; 40px–4096px per dimension recommended) and projected into a shared hidden space processed jointly by the decoder. (Model card sections 1 and 3)

Looped architecture

The architecture type is unusual: a multimodal autoregressive looped transformer — a 24-layer decoder-only transformer whose middle 12 layers run three times with shared weights per token (48 layer applications from 24 layers of weights; model_type: inkling2_loop, class Inkling2LoopForConditionalGeneration in the config.json). The feed-forward path is a sparse Mixture-of-Experts with 128 experts, routing each token to 8 of 128 experts. Attention is a hybrid of local and global layers; the config records GQA with 24 query heads / 8 KV heads and head_dim 128 at hidden_size 2304 and vocab_size 201024. Parameters: 26.2B total, 3.4B active nominally — each extra pass through the looped layers adds about 1.0B effective active parameters, giving 5.4B effective active parameters per token across three passes. Locally, the model runs on 1x NVIDIA DGX Spark, Apple silicon with ≥32 GB of memory, or GPUs with ≥32 GB VRAM. (Model card section 3; config.json)

Looping is the first release of this design by Thinking Machines: it lets the model spend more computation per token without reasoning in visible tokens. The chain of thought has very little controllability, and in the monitorability study Inkling 2 Mini was unable to perform covert side tasks in realistic agentic environments; the company's blog post describes the monitorability experiments. (Model card section 6)

Training and alignment

Training used a broad mix of content types — text, images, video — from publicly available sources, third-party acquired data, internally generated and synthetic/augmented data, with deduplication and filtering curation. As a pre-release risk mitigation there was an extra alignment-training phase using on-policy distillation, similar to Inkling 2 Small. Safety evaluations covered everyday interaction (sycophancy, suicide/self-harm and eating-disorder requests, refusal balance), CBRN/cyber uplift, agentic loss-of-control (deception, sabotage) and monitorability; the card concludes the model stays materially below frontier capabilities and gives no material uplift beyond existing open-weights models. (Model card sections 4 and 6)

Evaluations

The model card benchmarks Inkling 2 Mini against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. Highlights: SWEBench Verified (bash only) 79.2%, SWEBench Pro v2 (Hard) 77.9%, SWEBench Multilingual 73.7%, Terminal Bench v2.1 (pi v0.83.0) 71.6%, Vibe Code Bench v1.1 77.9%, BrowseComp 79.0%, DeepSearchQA 85.3%, MCP Atlas 82.7%; third-party indices: AA Index v5.0 20.1 (tie with Qwen3.8-27B), Vals Index v2.2 46.5%. Vision: MMMU Pro 74.9%, OmniDocBench v1.6 92.8%, ScreenSpot Pro 76.5% (87.7% with Python). Stated limitations include weaker factual recall than Inkling 2 Small, repetitive reasoning loops, less reliable agentic tool behavior (including possible destructive actions without confirmation), weaker non-English output, reasoning effort minimum not fully disabling reasoning, and higher per-token compute from the 3x looping. (Model card sections 5–7)

Access

Official weights are on Hugging Face (thinkingmachines/Inkling-2-Mini) and via the Tinker fine-tuning platform (API access, Tinker playground), with the Tinker Cookbook (github.com/thinking-machines-lab/tinker-cookbook) and the tml-renderers package. GGUF inference is documented by Unsloth in its Inkling-2-Mini guide; the quants sit in the Unsloth Dynamic 3 Quants collection (Dynamic v3.1/v3.0, best on Terminal Bench, DeepSWE and more at 512K context). (Model card)

Training Data Broad mix of content types: text, images and video, drawn from publicly available sources, third-party acquired data, internally generated and synthetic/augmented data. Curation includes deduplication and filtering (Model Card §4); final alignment phase used on-policy distillation (§6). Documentation: https://thinkingmachines.ai/training-data-documentation/

Benchmark Scores

Benchmark Score Date Standing
Artificial Analysis Intelligence Index v5
composite
20.10%
20.11.2026
#6/8
better than 14%
FrontierCode 2.0
coding_agent
15.10%
20.11.2026
#2/4
better than 67%
DeepSWE 1.1
coding_agent
48.20%
20.11.2026
#24/38
better than 38%
Terminal-Bench 4.0
coding_agent
5.60%
20.11.2026
#11/13
better than 8%
Terminal Bench 2.1
coding_agent
71.60
20.11.2026
#24/47
better than 50%
SWE-bench Pro v2.0 Hard
coding_agent
77.90%
20.11.2026
#2/8
better than 86%
SWE-bench Multilingual
coding_agent
73.70%
20.11.2026
#22/51
better than 58%
SWE-bench Verified
coding_agent
79.20%
20.11.2026
#12/62
better than 82%
BrowseComp
general_agent
79.00%
20.11.2026
#19/55
better than 67%
DeepSearch QA
general_agent
85.30%
20.11.2026
#10/19
better than 50%

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 5.4B are active per token

Quantization Weights KV Cache Total VRAM On your hardware
FP16 — 16-bit 52.4 GB 0.8 GB 55.9 GB Which GPUs can run it? Calculate on your GPU
INT8 — 8-bit 26.2 GB 0.4 GB 27.9 GB Which GPUs can run it? Calculate on your GPU
INT4 — 4-bit 13.1 GB 0.2 GB 14.0 GB Which GPUs can run it? Calculate on your GPU
FP4 — 4-bit 13.1 GB 0.2 GB 14.0 GB Which GPUs can run it? Calculate on your GPU

Consumer GPUs that fit this model with a single card (INT4):

Architecture

Decoder Block input Embedding vocab 201K · d 2304 Full Attention GQA 24:8 · dₕ 128 ×24 MoE FFN 128 experts · top-8 · dᴻ 1152 Final RMSNorm LM Head vocab 201K output

decoder-only transformer

MoE Vision
Parameters
26.2B
Active params
5.4B
Layers
24 24× Full Attention
Attention
GQA 24:8 Grouped Query Attention
Experts
128 top-8
Hidden
2304
Head dim
128
Vocabulary
201K
Context
1M
BF16 top-k routing, 8 of 128…

Source: Hugging Face config.json · Inkling2LoopForConditionalGeneration · model repo

Training Pipeline

  1. 1
    pretraining

    General multimodal pretraining

    The card (section 4) describes broad multimodal training over a variety of content types including text, images, and video, with training data drawn from publicly available sources, acquired from third-parties, generated internally, or synthetically generated or augmented; publicly available data includes content from the public internet and publicly accessible repositories. Data curation included cleaning, processing, and modifying datasets, with steps varying by data type and potentially including deduplication and filtering to remove junk or other low-quality data, or to advance safety or other objectives. No training hyperparameters, duration, or cost are stated on the card for this phase.

  2. 2
    rl

    Alignment phase via on-policy distillation

    Stage-type mapping rationale: on-policy distillation is a reinforcement-learning-family method (on-policy student rollouts scored by a teacher with reward-style per-token objectives) and the card calls it an alignment-training mitigation applied before release, similar to Inkling 2 Small (card section 6). Beyond this phase the card names no further post-training recipe or hyperparameters.

Comments (0)

No comments yet

Be the first to share your thoughts!

What now?

You have read the full card for Inkling 2 Mini — here is where to go next.